Anthropic will deploy up to 2GW of AMD AI infrastructure for training and running inference on Claude. AMD and Anthropic announced the deal on July 22, with the first 1GW beginning deployment in the first half of 2027. Anthropic's previously disclosed semiconductor infrastructure consisted of three platforms—AWS Trainium, Google TPU, and NVIDIA GPU—but this announcement adds AMD as a fourth. The relationship between the two companies doesn't start with a future large-scale procurement. Anthropic already uses the current-generation Instinct MI355X, and the collaboration is expanding to include deployment of next-generation hardware, joint software development, and an investment of up to $5 billion from AMD.
First 1GW to Deploy Starting First Half of 2027
The first 1GW will use the Helios rack-scale platform, which houses the MI455X from the Instinct MI450 series. The host CPU is the next-generation EPYC "Venice," networking is handled by AMD Pensando, and the software stack runs on ROCm. This isn't so much a contract to sell GPUs as a rack-level adoption that bundles compute, networking, and software together in AMD's design.
According to AMD's specifications, each Helios rack houses 72 MI455X units with a combined 31TB of HBM4. Compute performance is rated at up to 2.9 exaFLOPS for FP4 and up to 1.4 exaFLOPS for FP8. However, these are platform specifications based on AMD's internal analysis, not figures representing actual processing speed when running Claude.
What the announcement firmly commits to is that deployment of the first 1GW will begin in the first half of 2027. For any amount beyond the initial 1GW, neither whether it will be deployed nor when has been disclosed. The scale, use cases, and locations of the existing MI355X deployment also remain undisclosed. What's confirmed here is the deployment pathway that adds AMD and its upper limit—not a schedule for when the full 2GW will be operational.
From Three Platforms to Four: Claude's Procurement Strategy
As of May 2026, Anthropic had stated that Claude was trained and operated using AWS Trainium, Google TPU, and NVIDIA GPU. With AMD's addition, the semiconductor platforms confirmable through public information now number four. This isn't a switch from one vendor to another, but rather a further expansion of a procurement strategy that allocates workloads across multiple compute platforms.
| Semiconductor Platform | Publicly Disclosed Key Agreements | Deployment Timeline & Scale |
|---|---|---|
| AMD Instinct | MI450 series deployed via Helios | Up to 2GW, first 1GW deploying from H1 2027 |
| AWS Trainium | Securing new compute resources from Amazon | Up to 5GW, nearly 1GW by end of 2026 |
| Google TPU | Expanding next-gen TPUs with Google, Broadcom | 5GW, deployment starting from 2027 |
| NVIDIA GPU | Contract via Azure and use of SpaceX's Colossus 1 | Azure: $30 billion, up to 1GW. SpaceX: over 300MW, over 220,000 units |
These figures can't simply be added together, since they mix purchases with cloud usage, and contract ceilings, operational capacity, and deployment timelines all differ. Even so, the addition of AMD carries clear significance. Anthropic has stated a policy of allocating workloads to hardware suited to each use case, and this will allow real-world comparison of where the MI455X and ROCm can handle training versus inference.
Seller Becomes Buyer's Shareholder Through Up-to-$5-Billion Investment
AMD has committed to investing up to $5 billion in Anthropic in the future. This is a deal in which a semiconductor supplier also becomes a shareholder of its major customer, folding expanded adoption of the MI450 series and capital participation into the same collaboration. The announcement doesn't disclose the timing of the investment or the valuation. The equity stake and conditions for any staged execution also remain unknown, so the maximum amount and the actual amount invested need to be viewed separately.
In AMD's 6GW deal with OpenAI from October 2025, the direction of capital flow was reversed. AMD issued warrants for up to 160 million shares of its stock to OpenAI, structured so that the rights vest based on the initial 1GW, additional purchases, stock price, and technical and commercial conditions. In the announcement with Anthropic, it's AMD investing in the buyer. While both deals share the common thread of tying supplier and customer interests together through capital, AMD stock warrants and an investment in privately held Anthropic have different implications for AMD's finances and capital structure.
Implementation Conditions to Verify Before Performance
Helios isn't a finished rack that AMD sells directly—it's a reference design meant for OEMs and ODMs to productize. While AMD expects mass deployment to begin in the second half of 2026, it hasn't disclosed how many companies will handle manufacturing for Anthropic, or how many units will be installed at how many sites. Turning the 2GW ceiling into actual compute resources will require not just MI455X supply but also rack manufacturing capacity. On the data center side, power, cooling, and networking will also need to be prepared.
On the software side, the two companies are beginning a multi-year technical collaboration. This includes using Claude to optimize workloads for Instinct and accelerate ROCm development, as well as AMD's own plan to broadly deploy Claude across its design and product development divisions. This isn't an announcement of results achieved, but a development process aimed at running Claude efficiently on the MI455X. If the first 1GW deploys as scheduled and ROCm demonstrates a track record in training and inference operations, AMD could grow what started as a single large order into an ongoing compute foundation for Claude. What will determine success is whether deployment of the first 1GW in the first half of 2027 goes as planned, and whether additional capacity toward the 2GW ceiling actually materializes.